Moss - Real-time Semantic Search for Conversational AI

Every limit, error status and error code of Moss Inference, with what to do about each. Responses and Chat Completions share them.

Limits

Limits apply per project.

Limit	`moss/fastlane-1`	Vendor model
Requests per minute (sustained / burst)	120 / 10	300 / 50
Input tokens per minute	600,000	None
Prompt tokens per request	262,144	128,000 on the Moss key, 200,000 on your key
Indexes per request	32	32
Request body	4 MB	4 MB

On moss/fastlane-1, the 262,144 tokens are the whole context. The prompt, the retrieved facts and the reply share them. On the Moss key, anthropic/claude-haiku-5.5 takes up to 64,000 prompt tokens. Models lists the prompt limit of every model. One request searches at most 32 indexes, named in X-Moss-Indexes. MCP servers add limits of their own. A request with too many servers or tools returns 400. Too many MCP requests at once return 429 with code mcp_loops_busy. A request with MCP servers can run for a few minutes. To raise a limit for your project, ask through Support.

Rate limit headers

These headers show how much of the project’s limits is left.

Header	Value
`x-ratelimit-limit-requests`, `x-ratelimit-limit-tokens`	The limit
`x-ratelimit-remaining-requests`, `x-ratelimit-remaining-tokens`	What is left of it
`x-ratelimit-reset-requests`, `x-ratelimit-reset-tokens`	When it resets

The tokens headers come on moss/fastlane-1 only.

Errors

Every error has OpenAI’s shape: an error object with message, type and code.

429 Too Many Requests

{
  "error": {
    "message": "Rate limit reached for project 1a2b3c: too many requests per minute. Please try again in 1s.",
    "type": "requests",
    "code": "rate_limit_exceeded"
  }
}

Status	`type`	`code`	Cause	What to do
`400`	`invalid_request_error`		A malformed body, more than 32 index names, or a bad `tools` entry	Fix the request.
`400`	`invalid_request_error`	`context_length_exceeded`	A prompt over the model’s limit	Shorten the prompt. On `moss/fastlane-1`, leave room for the reply.
`400`	`invalid_request_error`	`model_not_enabled`	A listed model, on the Moss key, that is not served right now	Send `moss/fastlane-1`, or the model on your own key.
`400`	`invalid_request_error`	`missing_provider_key`	A vendor’s own model name, such as `gpt-6-luna`, without your key	Send your key. See Bring your own key.
`400`	`invalid_request_error`	`model_not_found`	A model name Moss does not serve	Pick an id from Models.
`400`	`invalid_request_error`	`unsupported_parameter`	A field or message part the model does not take on the Moss key	Remove it, or use your own key. See Models.
`400`	`model_error`		The model rejected the prompt	Change the prompt.
`400`	`mcp_error`	`mcp_server_error`	`server_url` is not a public `https` URL on port 443	Use a public `https` URL on port 443.
`401`	`authentication_error`		A missing or wrong `Authorization` token	Send `Authorization: Bearer <project_id>:<project_key>`.
`403`	`permission_error`		`X-Moss-Indexes` names an index outside this project	Name only this project’s indexes.
`413`	`invalid_request_error`		A body over 4 MB	Send a smaller body.
`424`	`mcp_error`	`mcp_server_error`	The MCP server did not connect, refused the credentials, or failed to list its tools	Check `server_url` and `headers`.
`424`		`failed_after_tool_call`	A failure after an MCP tool ran, which may have changed data	Do not replay it. See MCP servers.
`429`	`requests`, `tokens`	`rate_limit_exceeded`	Over a per-minute limit	Wait for `Retry-After`, then retry.
`429`	`usage_limit`	`usage_limit_exceeded`	Your credits ran out, so usage is paused	Add credits. See Pricing.
`429`	`capacity`	`mcp_loops_busy`	Too many MCP requests at once	Wait for `Retry-After`, then retry.
Vendor `4xx`	`model_error`	`upstream_error`	The vendor refused a call on your own key, with its own status	Read the vendor’s message, which follows `The model rejected the prompt`.
`502`	`model_error`		The model failed, or a vendor returned `5xx`	Retry with backoff.
`503`	`server_error`		Moss could not check the project	Retry with backoff.
`503`	`server_error`	`project_settings_unavailable`	Moss could not read the project’s system prompt, so it did not answer without it	Wait the 2 seconds of `Retry-After`, then retry.
`503`	`server_error`	`billing_unavailable`	Moss cannot meter a call on the Moss key right now	Wait the 5 seconds of `Retry-After`, or send `moss/fastlane-1`.
`503`	`capacity`	`server_overloaded`	Moss is at capacity, or the vendor rate limited the Moss key	Wait for `Retry-After`, then retry.
`503`	`capacity`	`vendor_unavailable`	The vendor refused or could not serve a call on the Moss key, such as on Anthropic’s `529` overload	Wait for `Retry-After`, then retry.

On your own key, a bad key returns the vendor’s status: 401 from OpenAI and Anthropic, 400 from Google.

Errors after a stream starts

A failure after a stream starts arrives as a stream event. On Responses, the stream ends with response.failed, whose response.error has type, code and message. On Chat Completions, a last data: frame carries error in place of choices, then data: [DONE] follows. Without stream, the same failure returns 502, or 424 once an MCP call ran. Streaming shows both stream forms.

Retry rules

The Moss SDKs retry on their own, 2 times by default. They retry a connection failure, or a 408, 409, 429 or 5xx, before the response starts. Each retry waits for Retry-After when it is 60 seconds or less, and backs off otherwise. In the Moss SDKs, x-should-retry: false stops a retry, and x-should-retry: true forces one. They never retry a stream that has started. The OpenAI SDKs retry on their own as well. If you call the endpoint directly, follow the same rules. Fix a 400, 401, 403 or 413 before you send the request again. Never replay a request that failed with failed_after_tool_call. An MCP tool already ran and may have changed data. The 424 carries x-should-retry: false, so the SDKs do not retry it either. Every response carries X-Moss-Request-Id, a refused one included. Include it when you report a problem. Each SDK exposes it as the request id of the response and of the error.

SDK errors

Each Moss SDK has one error for a failed call.

SDK	Error
Python	`moss.inference.InferenceError`
JavaScript	`InferenceError`
Browser	`InferenceError`
Swift	`InferenceError`
Android	`InferenceError`
Elixir	`Moss.Inference.Error`
C	A negative `MossResult`, with the fields in `moss_last_error_detail()`

The error holds the status, type and code from the Errors table. It also holds the request id and the Retry-After wait. JavaScript and the browser name the code errorCode. The SDK sets five types itself, with no status:

`type`	Cause
`invalid_request_error`	An argument the SDK rejects before sending, such as an empty key or an index name with a comma
`connection_error`	Moss Inference could not be reached, or the connection broke
`timeout`	A time limit ran out, such as the call’s timeout or idle timeout
`incomplete_stream`	The stream ended before `response.completed` or `response.incomplete`
`server_error`	The stream sent an event over 1 MiB, or one that is not a JSON object

Android throws IllegalArgumentException in place of invalid_request_error. In C, a rejected argument returns ERR_NULL_POINTER or ERR_INVALID_ARG, with a NULL detail.

Get started

Guides

SDKs

Reference

Errors and limits

Limits

Rate limit headers

Errors

Errors after a stream starts

Retry rules

SDK errors

​Limits

​Rate limit headers

​Errors

​Errors after a stream starts

​Retry rules

​SDK errors

Limits

Rate limit headers

Errors

Errors after a stream starts

Retry rules

SDK errors